The SIMA 2 agent from DeepMind is capable of playing any game

Google DeepMind introduced SIMA 2, a Gemini-powered AI agent that can understand instructions, reason, and teach itself new skills in virtual environments, doubling its predecessor's performance and nearing human-level task completion. The details: In previously unseen games like MineDojo and ASKA, the agent finished 45–75% of tasks, while SIMA 1 only finished 15–30% of the same tasks. SIMA 2 uses Gemini to construct tasks, assess attempts, and learn from mistakes in order to improve itself through trial and error without the need for human training data. By evaluating on-screen graphics, mimicking keyboard and mouse inputs, and engaging with the user like a gaming companion, the system navigates games. Additionally, DeepMind tested SIMA 2 in produced worlds from its Genie 3, where it was able to successfully adapt to training situations. SIMA 2 appears to be the largest step yet toward systems that can reason, communicate intelligently with people, and consistently conduct actions independent of the surroundings. Gaming is still a fantastic test ground for AI agents. A Gemini-powered agent might be our next in-game companion (or perhaps rival?).